App Disintermediation

The core claim

Peter Steinberger's thesis is that roughly 80 percent of consumer apps will disappear, because AI agents can manage data-centric tasks more naturally and conversationally than a purpose-built app ever could.¹ Most consumer apps exist to serve one narrow function, tracking fitness data, managing a to-do list, logging expenses, and their real value is a structured data model paired with a UI to interact with it. An agent with persistent, ambient access to a person's behavior can absorb all of it: it can already assume what someone ordering fast food is eating and log it without being asked, or simply be told to remind someone of something tomorrow without anyone caring where that reminder is stored. The apps expected to survive are the ones with an irreplaceable sensor or hardware layer, a camera-first app, a physical device controller, a medical device with a direct sensor connection, which is a small fraction of the total app ecosystem.

Where the value concentrates

If most apps disappear, Steinberger argues value concentrates in two places. Model companies supply the intelligence layer itself, and their advantage is structural because usage is sticky: complaints about token burn are, on his reading, evidence of attachment rather than frustration. The second place is the agent harness and its memory: whichever agent accumulates a person's history and preferences compounds that advantage over time, since the entity holding someone's memory effectively holds their attention. He also notes a counter-tension: models sit on a hedonic treadmill where each release feels transformative, quickly becomes the new baseline, and then draws complaints even though it has not gotten worse, and open source models keep closing the gap on closed ones. If models commoditize, the durable moat shifts entirely to the harness layer.

The interface counter-thesis

Brian Chesky agrees that apps give way to agents but rejects the assumption that the successor is a conversational text chatbot. For travel and e-commerce specifically, he argues a chatbot is the wrong interface because it is text-first, poor at comparison, and single-player, and that the actual successor is a rich, visual, directly manipulable, multiplayer surface, closer to a regenerated app than a message thread.² Chesky sharpens the question of where value lands, too: the guest-facing Airbnb app is only about 20 percent of the company, and the durable moat is the operation underneath it, payments, dispute resolution, the host network, none of which an agent can copy even once discovery itself is commoditized. So the two founders agree that agents beat standalone apps but disagree on the form factor of what replaces them.

Why it matters

The pattern reframes the calculation for a founder building a point app that mainly manages user data: that category of product is a declining proposition if an agent can subsume it, and the real strategic question becomes whether the app has a sensor, hardware, or network layer an agent cannot replicate. Several open questions remain unresolved, including whether people will trust a single general-purpose agent with the breadth of data a specialized app naturally compartmentalizes, and how app discovery works at all once agents, not app stores, are the primary interface.